Label emission is not moderation infrastructure

Labeler admissibility findings — operator-maturity scan, 2026-06-08

Summary

A scan of 150 observed ATProto labelers shows that label emission, declared consumer semantics, operational liveness, and moderation authority are separate properties. All three of the adjacent implication failures between them appear at population scale in the current Bluesky labeler ecosystem.

The central object:

Each row is a real-data finding with receipts; see the F-NNN entries below. The chain is one-directional adjacent non-entailment — not a claim that every pair of these properties has been separately disproven, but a claim that the adjacent implications cannot be silently assumed.

What was measured

For each labeler with either recent label-event activity (events_30d > 0) OR an ever-ingested service record on file, the scan computed:

• emission proxies: label_count_30d, label_count_7d, distinct_targets_30d
• declaration proxies: latest_label_def_count (count of labelValueDefinitions in the most recent observed service record), explains_labels (at least one definition has locale text), user_visible_consequence_known (declared with a defaultSetting in {hide, warn, ignore})
• liveness proxies: active_recently (any events in last 30d), has_stable_service_record (1 ≤ revisions ≤ 5), service_record_revisions total
• authority proxies: see the consumer-conversion census; here, only the is_reference and platform-root flags appear directly

Source: /var/lib/labelwatch/labelwatch.db on prod (192.46.223.21), scanned 2026-06-08 via docs/analysis/tools/operator_maturity_scan.py.

150 rows returned.

Core doctrine

The canonical anti-laundering hook (frozen in docs/specimens/DISAGREEMENTS.md):

Do not let "labeler exists" silently convert into "moderation service exists." Label emission, declared semantics, and operational liveness are separate properties. The observed ATProto labeler ecosystem contains all three failure modes at scale.

The three adjacent non-entailments, each receipted:

• emission ≠ declaration — see F-007 below
• declaration ≠ liveness — see F-008 below
• liveness ≠ authority — see F-001, F-004, and the consumer- conversion census

Longer-distance pairs (emission ≠ authority, declaration ≠ authority) are derivable from the chain but argued separately elsewhere; they are not direct outputs of this scan.

Operator-maturity scan

Class histogram (heuristic; see Limits):

abandoned                  65  (43%)
unknown                    26  (17%)
experimental               24  (16%)
community-service          13   (9%)
personal/reputational      11   (7%)
moderation-infrastructure  10   (7%)
platform-root               1   (<1%)
─────────────────────────────────
total                     150

The scanner sources, full per-row data, and reproduction instructions live in Receipts.

F-007 — emission without declaration

14 of 150 observed labelers emit at high volume (>1000 events in the last 30 days) while publishing zero labelValueDefinitions in their service record. Top of the cohort:

Two of these emit at moderation.bsky.app's own volume (~908k/30d each).

Why this matters. Under ATProto's stackable-moderation subscription model, client-side honoring of labels via labelersPref depends on a declared label vocabulary in the labeler's service record. A labeler with no labelValueDefinitions presents no standard consumer-side semantic surface to subscribe to, regardless of how many labels it emits.

Consumers, auditors, and downstream tools cannot infer that high-volume emitted labels correspond to an admissible moderation surface. Emission volume is therefore not evidence of usable protocol participation. A label-events firehose view of the ecosystem systematically overstates the size of subscribable moderation infrastructure.

Diagnostic class — four distinct gaps, none of which is "bug in labeler":

1. Protocol affordance gap — emitting labels is operationally easier than declaring usable consumer semantics.
2. Client reality gap — clients may rely on definitions; emitters may not provide them.
3. Ecosystem measurement gap — raw label volume overstates meaningful moderation infrastructure.
4. Anthropology gap — the "federated moderation ecosystem" is mostly not a mature operator field.

Generalization. F-007 generalizes F-006 (the moderation.bsky.app/needs-review specimen) from a single instance into an ecosystem pattern. Label publication is operationally separable from declared consumer semantics, and the separation occurs at non-trivial scale across many independent operators.

F-008 — declaration without liveness

65 of 150 observed labelers are abandoned (43%) — had a service record on file, zero events in the last 30 days.

Of those 65, 28 retain substantial declared scope (≥ 6 labelValueDefinitions) despite operational silence. Top of the substantial-scope subset:

Why this matters. A Bluesky user who subscribes to any of these labelers via labelersPref would receive zero label events from that subscription. The labeler is discoverable via its service record and may even appear in client UIs as available, but the subscription is operationally a no-op.

F-008 is the mirror image of F-007: declared consumer semantics without operational label emission. The same admissibility hook applies in both directions — declaration and liveness travel separately.

Pathological subcase — definition churn without emission. Some of the "abandoned" labelers are not "stopped" but "stuck":

vocalabeller.kanshen.click republished its service record 106,000 times with a single labelValueDefinition and zero events. The mirror of F-007's "shout into the void" is F-008's "republish without output." Both fail the consumer-protocol contract from different sides.

Limits / classifier caveats

This page is descriptive aggregate signal. Not measurement, not judgment, not accusation.

The maturity_class field is heuristic. Class boundaries are defensible but arbitrary (events_30d ≥ 100, events_30d ≥ 10000, etc.). Threshold tuning will move boundary cases. Do not cite an individual labeler's maturity_class as a normative judgment about that labeler. Cite the table as aggregate signal — the F-007 + F-008 cohort sizes and patterns. See D-002 for the discipline note.

has_contact_or_appeal_path is uniformly unknown_v1. Populating it would require appview profile fetches plus manual classification of description text; deferred.

Known classifier debt (T-001): the pre-existing labelers.likely_test_dev field flags xblock.aendra.dev (908k events/30d, 13 declared defs — the labeler our specimens-track cites as a real third-party adoption case) and recordcollector.edavis.dev (50k/30d, 67 defs) as test/dev. The maturity classifier uses likely_test_dev as an override and downgrades both to "experimental." The cohort framing in F-007/F-008 does not depend on these two specific classifications; the misflag is noise at the high-volume tail. See T-001 for the review item.

Sampling scope. Only labelers Labelwatch has discovered are in scope (labelers_total = 501 ever-discovered; 150 with recent activity or ingested service record). Labelers we have never observed are out of scope by construction. The consumer-conversion census also did not cover ~40 of ~47 clients in the Bluesky showcase; see that page for client-side sampling caveats.

Time window. Counts are "last 30 days" (events_30d) at scan time. Abandonment status is volatile — an "abandoned" labeler that resumes will move classes in the next scan. Regenerate the snapshot before citing these numbers in any time-sensitive context.

No per-account dossiers. Per labelwatch's CLAUDE.md: aggregate-first, descriptive language, no accusation-shaped outputs. This page operates per-labeler, not per-account.

Why this matters

The protocol model of stackable moderation imagines independent labelers publishing declared semantics, users opting in, and clients applying scoped moderation. The observed ecosystem, viewed through this scan, contains all three adjacent-implication failures of that model:

• High-volume labelers emit testimony their service records do not declare. Clients cannot subscribe to these labelers via the standard protocol mechanism — there is nothing to subscribe TO. (F-007)
• Labelers with rich declared scope go operationally silent. Users who subscribe to them get zero label events; the subscription is a no-op. (F-008)
• Even live, well-declared third-party labelers are not converted by default by any production client we sampled. Adoption is opt-in; the consumer-side conversion edge is not promoted from labeler publication. (consumer-conversion census + F-001 + F-004)

These are not bugs in any specific labeler. They are properties of the ecosystem at the current stage of its development. The shape matters most for downstream tools and policy discussions that silently equate "a labeler exists for X" with "moderation is happening to X." This page exists so the conflation has somewhere to be refused.

Receipts / generated data

This page's headline numbers are computed from a fixed snapshot; re-running the scanner against a fresh database may produce different cohort sizes (the ecosystem evolves).

Snapshot artifact (pinned by this page): - artifacts/operator-maturity-scan-2026-06-08.json — full per-labeler data, 150 rows × 23 columns + scan metadata - artifacts/operator-maturity-scan-2026-06-08.csv — flattened CSV for spreadsheet / research use - artifacts/operator-maturity-summary-2026-06-08.md — one-page TL;DR

Scanner source: - docs/analysis/tools/operator_maturity_scan.py — scan; reproduction: bash sudo -u labelwatch python3 operator_maturity_scan.py \ --db /var/lib/labelwatch/labelwatch.db \ --out /tmp/operator_maturity.json

Regression check (asserts the snapshot still produces the cited headline numbers — F-007 cohort = 14, F-008 abandoned = 65, F-008 substantial-scope subset = 28, vocalabeller specimen = 106k revisions / 1 def / 0 events, T-001 mis-flags present, full class histogram): - regression/test_findings_regression.py bash cd docs/findings/operator-maturity/ python3 regression/test_findings_regression.py # PASS (150 rows; all 7 regression checks green) This does NOT assert the numbers must stay stable forever — the point is that the SCANNER must explain when they change. If the regression fails, investigate why before updating the page.

Methodology + analysis writeup: - docs/analysis/labeler-operator-maturity-001.md — fuller analysis with all four notable patterns

Adjacent findings cited above: - F-001 — reference-labeler status does not imply default-client conversion - F-004 — third-party labelers publish labelValueDefinitions; opt-in consumers honor them - F-006 — needs-review is emitter-undeclared, not ingestion-gap-shaped (the F-007 single-specimen progenitor) - F-007 — population-scale publish-without-declare (this page) - F-008 — stale-service / abandoned declared-scope no-op subscriptions (this page) - D-002 — operator-maturity taxonomy is heuristic, not normative - T-001 — likely_test_dev mis-flags xblock + recordcollector

Consumer-side context: - docs/analysis/consumer-conversion-census.md — the "liveness ≠ authority" half of the chain. No sampled client hardcodes a non-mod.bsky labeler as a default.